Damage pattern mining in hurricane image databases

We present a damage pattern mining framework for hurricane data on residential houses using aerial photographs with 1:3000 based scale. The vertical photographs are normally collected for an overview of a disaster area or for more detailed assessment. Damage on roof, especially for shingles torn off, is expected to be discovered in a more efficient and automatic way instead of going through the high-resolution aerial photograph for numerous details. The discovered damage patterns can then be used for hurricane damage assessment and impacts on different geographical areas. Our methodology is: (1) applying a novel and effective segmentation method on each residential house on the aerial photograph of one community, (2) using the segmentation results to obtain a set of indexing parameters for each house representing the damage level of roof cladding as well as the patterns of damage, (3) using these parameters to select several templates representing the damage patterns so that users can issue query-by-example (QBE) queries. The proposed segmentation method is an unsupervised simultaneous partition and class parameter estimation algorithm that considers the problem of segmentation as a joint estimation of the partition and class parameter variables. By utilizing this segmentation method, the indexing parameters can be obtained automatically. The QBE capability can assist in finding similar damage patterns on the roof of the residential houses in different locations in the image databases. Experiments based on the aerial photographs of Hurricane Andrew in 1992 are conducted and analyzed to show the effectiveness of the proposed hurricane damage pattern mining framework.